Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Concept Learners for Few-Shot Learning
Kaidi Cao, Maria Brbic, Jure Leskovec
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| learn and aggregate concept-specific metric spaces and prototypes | few-shot classification · CUB, Tabula Muris, and Reuters | accuracy: COMET achieves 9.5% and 9.3% average improvements over the best performing baseline in the 1-shot and 5-shot tasks acros | Improved |
| use Euclidean rather than cosine distance (ablation) | COMET · CUB and Tabula Muris, 1-shot and 5-shot | accuracy: CUB: 65.7 ± 1.0 -> 67.9 ± 0.9; 82.2 ± 0.6 -> 85.3 ± 0.5; Tabula Muris: 77.1 ± 0.9 -> 79.4 ± 0.9; 90.1 ± 0.6 -> 91.7 ± 0. | Improved |
| share weights across concepts (ablation) | COMET · CUB, Tabula Muris, and Reuters, 1-shot and 5-shot | accuracy: Tabula Muris: 79.4 ± 0.9 -> 78.2 ± 1.0; 91.7 ± 0.5 -> 91.0 ± 0.5; Reuters: 71.5 ± 0.7 -> 69.8 ± 0.8; 89.8 ± 0.3 -> 88.6 | Mixed |
| use concept-specific learners rather than an ensemble of independent ProtoNets (ablation) | COMET · CUB, Tabula Muris, and Reuters, 1-shot and 5-shot | accuracy: CUB: 64.0 ± 0.8 -> 67.9 ± 0.9; 82.3 ± 0.5 -> 85.3 ± 0.5; Tabula Muris: 67.2 ± 0.8 -> 79.4 ± 0.9; 83.6 ± 0.5 -> 91.7 ± 0. | Improved |
| add concept dimensions beyond the whole-input concept (ablation) | COMET · CUB and Tabula Muris | accuracy: CUB: 10% and 5% with one beak concept for 1-shot and 5-shot; Tabula Muris: 7% and 17% with 8 concepts for 1-shot and 5-s | Improved |
| use automatically extracted concepts instead of human-defined concepts (ablation) | COMET · Tabula Muris and Reuters, 1-shot and 5-shot | accuracy: 2% lower accuracy on Tabula Muris and 1% on Reuters | Hurt |
| use automatically extracted visual concepts (ablation) | COMET · CUB and Flowers, 1-shot and 5-shot | accuracy: CUB: 64.8 ± 1.0, 82.0 ± 0.5; Flowers: 70.4 ± 0.9, 86.7 ± 0.6 | Improved |
| use multiple concepts rather than one whole-image concept (ablation) | COMET with Conv-6 · CUB, 1-shot and 5-shot | accuracy: 68.9 ± 0.9 -> 72.2 ± 0.9; 83.8 ± 0.6 -> 87.6 ± 0.5 | Improved |
| aggregate concept distances rather than use majority voting (ablation) | COMET · CUB, Tabula Muris, and Reuters, 1-shot and 5-shot | accuracy: CUB: 64.0 ± 0.9 -> 67.9 ± 0.9; 82.1 ± 0.6 -> 85.3 ± 0.5; Tabula Muris: 75.6 ± 0.8 -> 79.4 ± 0.9; 88.2 ± 0.5 -> 91.7 ± 0. | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DeepEMD — DeepEMD: Few-Shot Image Classification With Differentiable Earth Mover's Distance and Structured Classifiers (CVPR 2020) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Robust Meta-Representation Learning via Global Label Inference and Classification (TPAMI 2023 · similarity 0.89)
- Data-Centric Meta-Learning for Robust Few-Shot Generalization (CVPR 2026 · similarity 0.87)
- Bi-Level Meta-Learning for Few-Shot Domain Generalization (CVPR 2023 · similarity 0.87)
- Revisiting Unsupervised Meta-Learning via the Characteristics of Few-Shot Tasks (TPAMI 2022 · similarity 0.87)
- Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few Examples (ICLR 2020 · similarity 0.87)
- Unsupervised Meta-Learning via In-Context Learning (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
